library(pROC)
library(kernlab)


build_AUC_nonlinear_models = function(X, y, seed_value=456, build_mda_model=TRUE){
    #
    # Builds (using caret's train function) several of the nonlinear models discussed in
    # this chapter of the book optimizing the area-under-the-curve metric (AUC)
    #
    # X [n samples, n features] needs to have zero variance columns removed (use nearZeroVar to remove them) and
    # linear dependent columns removed (use findLinearCombos to remove them)
    #
    # y [n samples, 1] is a two factor vector of predictions with the FIRST factor corresponding to the event of interest
    #
    # Written by:
    # -- 
    # John L. Weatherwax                2009-04-21
    # 
    # email: wax@alum.mit.edu
    # 
    # Please send comments and especially bug reports to the
    # above email address.
    #
    #-----

    # Set up the train control arguments so that we can compute the area-under-the-curve:
    #
    ctrl = trainControl( summaryFunction=twoClassSummary, classProbs=TRUE )

    # Mixture Discriminant Analysis (MDA):
    #
    if( build_mda_model ){
        set.seed(seed_value)
        mda.classifier = train( X, y, method="mda", tuneGrid=expand.grid(subclasses=1:3), metric="ROC", trControl=ctrl )
        mda.predictions = predict( mda.classifier, X, type="prob" )
        mda.rocCurve = pROC::roc( response=y, predictor=mda.predictions[,1] )
        mda.auc = mda.rocCurve$auc[1]
        mda=list( classifier=mda.classifier, roc=mda.rocCurve, auc=mda.auc )
    }

    # Neural Networks:
    #
    set.seed(seed_value)
    nnetGrid = expand.grid( size=1:3, decay=c(0,0.1,1,2) )
    nnet.classifier = train( X, y, method="nnet", preProc=c("center","scale","spatialSign"), tuneGrid=nnetGrid, metric="ROC", trace=FALSE, maxit=2000, trControl=ctrl )
    nnet.predictions = predict( nnet.classifier, X, type="prob" ) # <- returns probability of "Yes" (event of interest) & "No"
    nnet.rocCurve = pROC::roc( response=y, predictor=nnet.predictions[,1] )
    nnet.auc = nnet.rocCurve$auc[1]
    nnet=list( classifier=nnet.classifier, roc=nnet.rocCurve, auc=nnet.auc )

    # Support Vector Machines:
    #
    set.seed(seed_value)
    sigmaEst = kernlab::sigest( as.matrix(X) )
    svarid = expand.grid(sigma=sigmaEst[1], C=2^seq(-4,+4))
    svm.classifier = train( X, y, method="svmRadial", tuneGrid=svarid, preProc=c("center","scale"), metric="ROC", fit=FALSE, trControl=ctrl )
    svm.predictions = predict( svm.classifier, X, type="prob" )
    svm.rocCurve = pROC::roc( response=y, predictor=svm.predictions[,1] )
    svm.auc = svm.rocCurve$auc[1]
    svm=list( classifier=svm.classifier, roc=svm.rocCurve, auc=svm.auc )

    # K-Nearest Neighbors:
    #
    set.seed(seed_value)
    knn.classifier = train( X, y, method="knn", tuneLength=20, preProc=c("center","scale"), metric="ROC", trControl=ctrl )
    knn.predictions = predict( knn.classifier, X, type="prob" )
    knn.rocCurve = pROC::roc( response=y, predictor=knn.predictions[,1] )
    knn.auc = knn.rocCurve$auc[1]
    knn=list( classifier=knn.classifier, roc=knn.rocCurve, auc=knn.auc )

    # Naive Bayes:
    #
    set.seed(seed_value)
    nb.classifier = train( X, y, method="nb", metric="ROC", trControl=ctrl ) 
    nb.predictions = predict( nb.classifier, X, type="prob" )
    nb.rocCurve = pROC::roc( response=y, predictor=nb.predictions[,1] )
    nb.auc = nb.rocCurve$auc[1]
    nb=list( classifier=nb.classifier, roc=nb.rocCurve, auc=nb.auc )

    result = list( nnet=nnet, svm=svm, knn=knn, nb=nb )
    if( build_mda_model ){ result = c(result, list(mda=mda)) }
    return( result )

}